About spa_sync() and zil_commit() execution synchronization

[dgiantsidi]

Hi there,

In my understanding on the codebase, the spa_sync() is not synchronized with the zil_commit(). Basically, that means that while the spa_sync() updates the new zil_header the tail of the ZIL can be written without blocking. As such, it is not guaranteed that the zil_header will always point to a HOLE, meaning that the entire ZIL (up to this point) has been trimmed.

Edit: Actually I see that the zil_sync that updates the new zil_header is synchronized with the updates on the zil with zl_lock.. However, my question still holds...

Do we always know that spa_sync() will point to a zil_header that is a HOLE?

[amotin]

May be I miss something, but why do you think zil_header should point to a HOLE after spa_sync()? There might be records in ZIL up to two TXGs ahead of one spa_sync() is syncing right now, and they should remain there in case we crash.